Actions
Engagement
  • Author Author: tariq.ahmad
  • Date Created: Date Created
  • Last Updated Last Updated:
  • Views 5616 views
  • Likes 12 likes
  • Comments 57 comments
Related
Recommended

Project14 | Vision Thing: Beaglebone AI Your Vision Thing Project!

image

Vision Thing

Enter Your Project for a chance to win an Oscilloscope Grand Prize Package for the Most Creative Vision Thing Project!

Back to The Project14 homepage image

Project14 Home
Monthly Themes
Monthly Theme Poll

 

 

In the Comments Below: Tell Us How You Would Use the BeagleBone AI for Your Vision Thing Project!

 

We'll Send Out Boards for Project Proposals that Use Them! 

 

 

image

 

 

We are offering up to 20 FREE Beaglebone AI Boards in exchange for Vision Thing projects that use them!

 

Beaglebone AI Cooling Cape Addon Available from :  BB AI cooling Addon board available

image
Buy NowBuy Now
We are offering up to 20 FREE Beaglebone AI Boards in exchange for Vision Thing projects that use them!

 

What is a Vision Thing project and how do you use the BeagleBone AI to do one?

 

There's a lot of variety with how you choose to implement your project.  It's a great opportunity to do something creative that stretches the imagination of what hardware can do.  Your project can be either a vision based project involving anything that is related to Computer Vision and Machine Learning , Camera Vision and AI based projects, Deep Learning, using hardware.  Or, it can be a graphics project involving something graphical such as adding a graphical display to a microcontroller, image processing on a microcontroller, image recognition interface a camera to a microcontroller,  or FPGA - camera interfacing/image processing/graphical display.

 

What makes the Beaglebone AI suitable for Vision Thing  Projects?

 

BeagleBone AI is a high-end board for developers building artificial-intelligence and computer-vision applications. Its main AI features include a Texas Instruments (TI) AM5729 system on chip (SoC), TI C66x digital-signal-processor (DSP) cores and embedded-vision-engine (EVE) cores. The board has the same form-factor as the popular and cheap BeagleBone Black but with much higher specifications.

 

 

The AI-ready board comes with 1GB RAM and 16GB on-board eMMC flash with a high-speed interface, a USB Type-C port for power and a dual-role controller, and a USB Type-A host. There's also Gigabit Ethernet and Wi-Fi. With preinstalled software, the BeagleBone AI also saves buyers from having to download equipment to get the device up and running.

image
The Most Creative Vision Thing Wins a Keysight DSOXO11G Oscilloscope!

Learn more by visiting:

So: what do you have to do to be in the running for one of the brand new Beagleboard AI? Just follow the instructions below!

 

1. Register and/or log-in to the element14 Community

2. Leave a comment on this post or on Project14 | Vision Thing: Build Things Using Graphics, AI, Computer Vision, & Beyond!  telling us what you'd like to build for your Vision Thing project with the new Beagleboard AI!

3.  Once you receive a new board submit your finished Vision Thing  project for a chance to win a Keysight DSOXO11G Oscilloscope!

 

 

The Most Promising Vision Thing Project Proposals Win a Free Beaglebone AI to Use In Your Vision Thing Project!

 

Submit Your Completed Project in Vision Thing  for a Chance to Win a Keysight DSOXO11G Oscilloscope!

Parents

  • Having more-or-less decided on something for the Vision Thing based on Arduino and possibly Processing as well, I have just seen this opportunity to use the BeagleBone AI. Being mostly used to using Arduino Nanos and the odd MAKERZERO and similar the BeagleBone AI seems to bucket loads of processing and memory and connectivity, it must need it's own power station if they are all used simultaneously! To be honest, I'm not sure that my Vision Thing idea needs all this processing power, just the FPU and large memory, but if I get one I'll see what else I could use on it.

     

    My idea is to make my own 'camera' out of 64 small LDRs arranged into an 8x8 grid array, connecting to some processing unit such as a MKRZERO (as I have one) or maybe a BeagleBone AI (if I get one). Analogue multiplexers would be used to connect to the ADC inputs available to obtain a grey scale image. Each bit of the image would be connected to an input on an Artificial Neural Network (ANN). For a simpler feed forward ANN this would require 64 neurons in the input layer, probably one hidden layer of something like 128 neurons, plus the output layer where the number of neurons is determined by the number of classes or types to be recognised. My initial activity will be to try and train the ANN to recognise the 26 letters of the alphabet. I would also like to implement some sort of graphical display to show the current image being obtained from the LDR camera. Initially I might try to use the LDR camera without a lense with just the letter being placed directly onto the surface of the LDRs, but if I can work out some lense stuff (that's a technical name used by anyone who doesn't know anything about optics) I might even try to make a proper camera out of it and obtain images of the 'world' . Then it would be possible to try and recognise a much wider range of classes or types. I could just use an existing camera, but where would be the fun in that!

     

    The main reason for using something like the BeagelBone AI is that an artificial neuron is essential a sum of products mathematical operation, so for example each neuron on the input layer, would have 64 inputs (one from each LDR) multiplied by an individual weighting constant (both of which work best as floating point values) which are then all added together, so another 64 addition operations. This is then replicated on each of the 64 input neurons (so 64 x 64 multiplies and so on). These values are then passed to the hidden layer where the number increases to 128 x (64x64) multiples, and then again to the output layer of 26 x (128 x (64x64) = 13,631,488 plus all the additions. So over 13 million floating point operations and this is why many ANN systems do not use such large numbers of neurons. Plus you need all the memory to store all the values. If I used a MKRZERO I would probably have to reduce the number of neurons in the hidden layer and possibly the output layer as well so only recognising a handful of letters, but maybe the BeagleBone AI can perform a brute force and hence quicker, solution. I'm not an expert on any of this so I might be completely wrong.

     

    It might all collapse and die at the first hurdle but I'm ready to give it a go.

     

    Dubbie

  • in reply to dubbie

    Is there any particular reason to not use semiconductors?

    I think one of the simplest approaches would be to use a pinhole, you definitely have to be very careful with light leaking in, or you won't be able to get much meaningful data. Also its very likely that you will need to calibrate the system, as each photosensor will respond different to light.

Comment
  • in reply to dubbie

    Is there any particular reason to not use semiconductors?

    I think one of the simplest approaches would be to use a pinhole, you definitely have to be very careful with light leaking in, or you won't be able to get much meaningful data. Also its very likely that you will need to calibrate the system, as each photosensor will respond different to light.

Children
  • in reply to neuromodulator

    I'm not sure whether to try and make some sort of pinhole camera, or lenses based camera, or maybe just stick with putting shapers over the LDR array to create the images.

     

    I am going to try and reduce light leakage by putting each LDR into a hole drilled into some wood which should eliminate any light coming in from the sides. Not sure how deep to make the holes though.

     

    There could well be a problem with different LDRs having different responses to the same amount of light. From the experiment with the 1x8 array, this seems to be minimal. If it is a problem I would be able to calibrate each LDR individually which would be a pain, but it would be achievable. I will be using an ANN to classify images anyway, so small differences in images caused by LDRs having different characteristics should not be too much of a problem.

     

    Dubbie